Project 1: SAR → EO Image Translation Using CycleGAN

---

1. Project Overview

We implement a CycleGAN to translate Sentinel-1 SAR imagery (2 channels) into Sentinel-2 EO imagery (3 channels).

Objectives:

• Normalize and preprocess SAR & EO patches
• Train CycleGAN under the SAR→EO (RGB) configuration
• Evaluate results with spectral SSIM, PSNR, and visual comparisons

---

2. Repository Structure

project/
├── Project1_SAR_to_EO/
│   ├── data/
│   │   ├── raw/            # original .pt samples  
│   │   └── processed/      # normalized train/val splits  
│   ├── checkpoints/        # saved model weights  
│   ├── generated_samples/  # output EO images & comparisons  
│   ├── preprocess.py       # preprocessing & splitting  
│   ├── train_cycleGAN.py   # training script  
│   ├── evaluate_results.py # metrics & visualization  
│   └── config.yaml         # hyperparameters & paths  
├── README.md               
└── requirements.txt        

---

3. Setup & Installation

1. Clone the repo:

   git clone https://github.com/your‐org/your‐repo.git
   cd your‐repo/Project1_SAR_to_EO
   

2. Create a virtual environment and install dependencies:

   python3 -m venv venv
   source venv/bin/activate
   pip install --upgrade pip
   pip install -r requirements.txt
   

---

4. Data Preprocessing

1. Place your 6,000 .pt files in data/raw/. Each file must be a dict:

   {'A': Tensor(2×256×256),   # SAR (VV, VH)
    'B': Tensor(3×256×256)}   # EO (RGB)

2. Run preprocessing and train/val split (80/20):

   python preprocess.py
   

   • Normalizes each channel to [–1,1]
   • Saves to data/processed/train/ and data/processed/val/

---

5. Training

Edit config.yaml to adjust hyperparameters (batch size, epochs, learning rate). Then run:

python train_cycleGAN.py

— Model checkpoints will be saved under checkpoints/ as G_A2B_epoch{n}.pth.

---

6. Evaluation & Visualizations

Run metrics computation and generate example images:

python evaluate_results.py

This will:

• Compute SSIM & PSNR for N validation samples
• Save comparison grids to generated_samples/

6.1 Suggested Visuals

Leave space below to paste or link your visuals:

1. Figure 1. SAR → EO Example (Before / After)

2. Figure 2. Additional Examples
3. Figure 3. Quantitative Metrics Over Samples

📊 Evaluation Metrics (after 30 epochs)

Sample	SSIM	PSNR	NDVI
sample_0000.pt	0.7110	24.58	0.5067
sample_0001.pt	0.6143	17.03	0.4683
sample_0002.pt	0.7328	26.68	0.4780
sample_0003.pt	0.6836	21.49	0.5638
sample_0004.pt	0.6749	26.36	0.5649
sample_0005.pt	0.7644	25.68	0.5153
sample_0006.pt	0.7750	26.93	0.5481
sample_0007.pt	0.6369	19.69	0.5864
sample_0008.pt	0.6452	19.65	0.6178
sample_0009.pt	0.7290	21.83	0.6507
sample_0010.pt	0.7109	22.55	0.6260

✅ Overall Performance:

• Average SSIM: 0.6980
• Average PSNR: 22.95 dB
• Average NDVI: 0.5569

4. **Figure 4. Training Curves **

---

7. Key Findings & Observations

• Normalizing each channel independently to [–1,1] stabilized training.
• Achieved average SSIM ≈ 0.69 and PSNR ≈ 22.95 dB on RGB.
• Failure modes: difficulty reconstructing fine texture in heavily vegetated zones.

---

8. Tools and Frameworks

• Python 3.8+
• PyTorch
• torchvision
• scikit-image (SSIM, PSNR)
• matplotlib

---

9. Next Steps (Future Work)

• Experiment with perceptual (VGG) loss for sharper textures
• Extend to NIR or SWIR target bands
• Integrate cloud-mask guided loss for cloudy regions

---

10. How to Reproduce

1. Ensure data in data/raw/
2. Install requirements and follow Sections 4–6 above

---